Pith. sign in

REVIEW 3 major objections 5 minor 60 references

Positive-Unlabeled Preference Optimization For Chest X-ray Report Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read PU-DPO corrects preference optimization for genuinely present but unmentioned chest X-ray findings.

desk verdict Solid method, clean semi-synthetic proof, but the real-world gains rest on a test-set-derived alpha—and the abstract overstates them. read the letter →

arxiv 2608.05341 v1 pith:7ZZLDDBU submitted 2026-08-05 cs.CV cs.LG

Yuta Kobayashi , Pradyun Ramesh , Muhammad Ahmed Chaudhry , Vincent Jeanselme , Judy Wawira Gichoya , Sanmi Koyejo , Kathleen Capaccione , Shalmali Joshi This is my paper · ORCID
classification cs.CVcs.LG
keywords positive-unlabeledlearningpreferenceoptimizationdirectchestX-rayreportgenerationomissionnoiseSCARassumptionvision-languagemodelshiddenpositives
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that standard preference optimization for chest X-ray report generation is quietly corrupted by omission noise: retrospective reports often leave visually present findings unmentioned, so a model trained to prefer reports matching the reference learns to under-report. PU-DPO reformulates the direct preference optimization objective as positive-unlabeled learning, treating explicitly mentioned findings as positive and absent mentions as unlabeled, and subtracts a contamination term $\alpha(R_p^+ - R_p^-)$ from the marginal loss $R_X^-$ so that hidden positives are not punished. The paper proves this estimator is unbiased for the true preference loss under the SCAR assumption and a known class prior, and shows in semi-synthetic and real-world experiments that it preserves sensitivity as omission noise rises while naive DPO degrades. A sympathetic reader would care because retrospective clinical reports are the only scalable supervision for radiology VLMs, and this is a direct correction to their systematic incompleteness.

What carries the argument

The central object is the unbiased PU-DPO loss $R_j(\phi) = \alpha_j R_{p,j}^+ - \alpha_j R_{p,j}^- + R_{X,j}^-$ for a disease subtype $j$, where $R_{p,j}^+$ is the expected log-loss for preferring the mention variant on true positives, $R_{p,j}^-$ flips that preference on true positives, and $R_{X,j}^-$ is the expected log-loss for preferring the omission variant over the full marginal of images. The mixture identity $P_X = \alpha P_p + (1-\alpha) P_n$ splits the marginal risk so that the unknown true-negative risk cancels algebraically, leaving only quantities computable from labeled positives and the unlabeled marginal, and the class prior $\alpha$ acts as a calibration dial that controls how strongly the model resists treating unmentioned findings as negative. Preference pairs are generated by an editor LLM that inserts or removes a single pathology subtype, keeping the comparison axis homogeneous within a batch so the same $\alpha_j$ applies.

What would settle it

Construct a dataset in which omission probability depends on an image feature (e.g., small effusions omitted more often than large ones) while holding the class prior fixed; if PU-DPO's hidden-positive recovery then falls to the level of naive DPO, the practical value of the unbiasedness guarantee is refuted. The paper's own cardiomegaly result, where a kernel MMD test rejects SCAR and the correction yields no gain, is a natural instance of exactly this test.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the preference signal can be de-noised of its dominant systematic error by carrying a PU-learning identity into the DPO objective: write the supervised loss as $\alpha R_p^+ + (1-\alpha) R_n^-$, replace the unobservable negative-side risk by the marginal risk $R_X^-$ minus $\alpha R_p^-$, and optimize $\alpha(R_p^+ - R_p^-) + R_X^-$. This is an unbiased estimate of the true loss under SCAR, and because hidden positives contribute to $\sigma(z_i)\nabla z_i$ in both PU-DPO and naive DPO, the difference is that PU-DPO's labeled-positive correction term $-(\alpha/n_p)\nabla z_i$ counters the wrong-direction push instead of letting it accumulate. The paper also constructs preference pairs by editing a sampled model response into variants that mention or omit a single target subtype, so the comparison is on a single clinical axis and the per-subtype PU correction applies. Empirically, the paper reports that PU-DPO maintains sensitivity near the noiseless oracle across simulated omission rates up to 30%, and on the adjudicated CheXpert set improves omission recovery for atelectasis and pleural effusion relative to the same DPO pipeline with minimal report-quality cost.

Load-bearing premise

The correction is unbiased only if, within each disease subtype, the probability that a true positive is mentioned in the report is the same for every image; the paper's own test shows this fails for cardiomegaly, where PU-DPO then fails to improve.

Editorial extensions

If this is right

  • As omission noise rises to 30%, PU-DPO keeps sensitivity near the noiseless oracle for pleural effusion (F1 around 0.84), while naive DPO's sensitivity drops by roughly 0.06 to 0.07.
  • On the CheXpert radiologist-adjudicated set, PU-DPO increases omission recovery over DPO for atelectasis and pleural effusion while specificity stays within about one point of the base model.
  • The method transfers across LLaVA-Rad, MedGemma, and a chain-of-thought model, though the CoT variant sees little benefit, which the paper attributes to limits in the editor's handling of long reasoning traces.
  • Underestimating the class prior collapses PU-DPO toward naive DPO behavior, so the paper's practical recommendation is to err on the high side when estimating $\alpha$.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The per-subtype decomposition suggests PU-DPO effectively recalibrates a per-finding diagnostic threshold, and the same loss could be applied to any structured output task whose reference labels are one-sided incomplete, such as discharge summaries or pathology reports.
  • The cardiomegaly failure indicates that a cheap SCAR diagnostic (like the paper's own MMD test) should precede deployment on a new pathology, and the natural extension is a version that models subtype-dependent labeling probabilities instead of assuming a constant rate.
  • Because contrastive pairs come from editing model outputs rather than human collection, PU-DPO may offer a low-cost route to noise-robust alignment in other domains where pairwise preferences are scarce but nominal labels are plentiful.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces PU-DPO, a direct preference optimization variant for chest X-ray report generation that treats absent mentions of findings in clinical reports as unlabeled rather than as true negatives. The method constructs contrastive response pairs by editing model-generated reports to include or omit a specific pathology subtype, and then optimizes the standard DPO objective with a positive-unlabeled risk correction scaled by the positive class prior alpha. The authors evaluate PU-DPO in semi-synthetic experiments with simulated omission noise, in a radiologist study quantifying omission rates in CheXpert, and on real-world CheXpert and MIMIC-CXR benchmarks using three VLM backbones. The headline claim is that PU-DPO recovers hidden positives and yields consistent detection-rate gains across pathologies compared to naive DPO and other baselines.

Significance. If the central claim were fully supported, PU-DPO would be a useful contribution: treating report omissions as unlabeled rather than negative is a natural and practically motivated correction, and the contrastive-edit preference construction is a clean way to obtain per-finding supervision. The semi-synthetic experiments are a genuine strength: they isolate the effect of the PU correction under controlled noise, include a standard DPO and a robust DPO baseline, and report bootstrap confidence intervals. The paper also provides reproducibility details, code, and a radiologist audit of the generated preference pairs, which is commendable. However, the real-world evaluation has a load-bearing circularity issue involving the class prior alpha, and the paper's own robustness analysis contradicts the unqualified 'consistent gains across multiple pathologies' claim. These issues do not invalidate the theoretical derivation, which is standard PU risk reweighting, but they do prevent the current manuscript from supporting the abstract's central empirical assertion.

major comments (3)
  1. [5.3, Figures 3-4, Appendix E.6] The main real-world evidence for omission recovery is not currently established because the manuscript does not disclose which alpha estimator was used to train the PU-DPO models reported in Figures 3 and 4. Section 5.3 defines both a data-driven alpha_BBE and alpha_adjudicated that 'has access to the oracle using the adjudicated test set,' and the sensitivity analysis in Appendix E.6 perturbs 'the adjudicated alpha estimate.' Since the paper states in Section 5.3 that alpha acts as a calibration factor controlling the sensitivity-specificity trade-off, training with alpha_adjudicated and then measuring omission recovery on the same adjudicated test set can manufacture gains that would not be obtained with a deployable pipeline. This concern is reinforced by Table 13, where overestimating alpha in the semi-synthetic setting improves F1 (from 0.84 to 0.86) by trading specificity. The abstract's claim of consistent real-world gains must therefore be supported by results using only data-driven alpha_BBE, with explicit reporting of the alpha value used for every model and figure; if alpha_adjudicated is intended only as an oracle sensitivity check, those results must be clearly separated from the headline conclusions.
  2. [5.3, Table 18, Figures 12-14, Abstract] The paper's own robustness analysis contradicts the unqualified claim of 'consistent gains across multiple pathologies.' Table 18 reports a statistically significant kernel MMD test for cardiomegaly (MMD p = 0.016), rejecting the SCAR assumption for that finding, and Appendix E.6 states that for cardiomegaly PU-DPO does not improve omission recovery over standard DPO. Since the abstract and Section 6 claim consistent improvements across multiple pathologies, the abstract should be revised to reflect that the real-world benefit is demonstrated mainly for atelectasis and pleural effusion, and that PU-DPO can fail when the multi-SCAR assumption is violated. The current framing of this limitation appears only in the appendix and the discussion, while the headline claim remains unconditional.
  3. [4.2, Eq. (4), Appendix B.6] The unbiasedness result in Proposition A.1 is stated for a single finding under the natural marginal distribution P_X, but the training procedure in Section B.6 performs stratified 50/50 sampling of labeled positives and unlabeled examples and samples subtypes uniformly. Because the empirical loss is an expectation over this modified batch distribution rather than over P_X, the unbiasedness guarantee does not directly apply to the actually optimized objective. The manuscript notes this distributional shift and says no practical issues were observed, but it does not provide a formal argument or an ablation isolating this effect. A simple experiment comparing balanced and naturally distributed batches, or a statement of the conditions under which the unbiasedness transfers to the stratified distribution, would make the theoretical claim match the implementation.
minor comments (5)
  1. [Section 1] The phrase 'anatomical loanatomical location' contains a typo and should read 'anatomical location.'
  2. [Appendix E.4, Table 10 caption] The caption states 'comparing PU-DPO with naive labels against PU-DPO'; this appears to be a typo, and it should likely read 'comparing naive DPO with PU-DPO.'
  3. [Appendix E.6, Figures 15-17] The sensitivity analysis is described as perturbing 'the adjudicated alpha estimate,' but the figure captions do not state which alpha estimator was used for the central point. Please label each figure with the estimator and the perturbation range explicitly.
  4. [Section 5.3 and Appendix B.6] The definition of alpha is confusing because Eq. (2) uses alpha = P(Y = +1), while the BBE procedure estimates P(Y = 1 | S = -1) and the batch-level alpha is then computed as (np + nu P(Y=1|S=-1))/n. This distinction should be explained in the main text rather than only in the appendix.
  5. [Table 1] The omission and hidden positive rates are based on small denominators (e.g., 14/32 for cardiomegaly); reporting confidence intervals for these proportions would help readers calibrate the strength of the radiologist-study evidence.

Circularity Check

1 steps flagged · score 6.0 of 10

Real-world omission-recovery gains are partially forced: the reported PU-DPO models appear to use α estimated from the same adjudicated test set that defines the headline metric, and α is acknowledged to control the sensitivity-specificity tradeoff.

  1. fitted input called prediction [Section 5.3, 'α-Estimation' and 'Omission recovery in adjudicated test set']
    "Second, we generate an estimate αadjudicated that has access to the oracle using the adjudicated test set, computing the empirical frequency of hidden positives to obtain pathology-level estimates of P(Ypathology = +1|S=−1)... On the adjudicated test set we compute omission recovery, the rate at which a model reports a finding that the original clinical report omitted but the adjudicating radiologists confirmed as present... Because PU-DPO is exactly our DPO baseline plus a α-scaled PU loss correction term, the choice of α effectively acts as a calibration factor."

    The headline real-world result is omission recovery on the radiologist-adjudicated CheXpert test set. The α used to train the reported PU-DPO models is evidently αadjudicated, estimated from that same test set's hidden positives, rather than the data-driven αBBE. Since the paper states that α is a calibration factor that directly controls the sensitivity-specificity tradeoff, setting α from the evaluation set's hidden-positive frequency and then measuring omission recovery on the same set is not an independent prediction: it is an oracle-calibrated knob being evaluated on its own calibration data. The α-sensitivity analysis perturbs 'the adjudicated α estimate,' confirming that the main real-world results are benchmarked with test-set oracle information.

full rationale

The core PU-DPO objective derivation (Section 4.1, Appendix A.1) is the standard PU unbiased-risk algebra: R = αRp+ − αRp− + RX−, obtained from the mixture identity PX = αPp + (1−α)Pn. This is self-contained and not equation-level circular; it relies on external PU-learning citations (Elkan and Noto, Du Plessis et al., Kiryo et al.) rather than self-citations. The semi-synthetic experiments use a known oracle α=0.5 and test robustness under controlled misspecification, which is a legitimate controlled evaluation. The circularity burden lies in the real-world evaluation: the paper introduces αadjudicated computed from the adjudicated test set, and the headline omission-recovery metric is defined on that same adjudicated set. Because α is explicitly described as a calibration factor controlling sensitivity-specificity, training with test-set-derived α and then reporting recovery of hidden positives on that test set is partially forced by construction. This is compounded by the paper's own caveats: it states that 'given the sample size we cannot yet conclude with statistical certainty' and that for cardiomegaly, which violates SCAR, 'omission recovery failing to improve over the naive DPO baseline,' so 'consistent gains across multiple pathologies' is also internally qualified. These limitations are honest but do not remove the oracle-α circularity. Score 6 reflects partial circularity: the method's independent content (PU risk correction, semi-synthetic validation) is substantial, but the central real-world predictive claim is not established without data-driven α results.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The method rests on standard PU mixture identities plus strong domain assumptions about how omission noise arises and how the editor LLM constructs pairs. No new physical or biological entities are introduced. The class prior alpha is the key free parameter and largely controls the reported behavior.

free parameters (3)
  • class prior alpha_j (per subtype) = 0.5 in semi-synthetic; BBE estimates 0.087 to 0.169 in real data; adjudicated-estimates in Appendix E.3
    The PU correction term is scaled by alpha. The paper's sensitivity analysis shows alpha directly controls the sensitivity-specificity tradeoff, and underestimating it collapses PU-DPO toward standard DPO.
  • DPO beta = 0.05
    Standard preference scaling parameter; not central to the PU claim.
  • length filter threshold tau = 0.5
    Filters contrastive edits whose length differs too much from the base response; a practical design choice.
assumptions (6)
  • domain assumption SCAR holds within each pathology subtype: labeling probability c is constant and independent of image features.
    Required for the unbiasedness of the PU estimator. The paper's own MMD test (Table 18) rejects this for cardiomegaly, and PU-DPO fails to improve omission recovery for that condition.
  • domain assumption Explicitly mentioned findings are true positives; there are no false-positive labeling errors.
    Section 3.1 states 'When si = +1, we assume that it is present.' This is load-bearing because labeled positives anchor the PU correction.
  • domain assumption The true preference for presence vs absence is determined solely by the true finding label Y under the Bradley-Terry model.
    Section 4.1 defines ground-truth preference by Yj. If a positive response is clinically wrong, the preference signal is corrupted.
  • domain assumption The editor LLM produces contrastive edits that accurately insert or remove the target finding while preserving unrelated content.
    Preference dataset construction in Section 4.2 depends on this. The audit in Appendix D qualifies it: only 70% of chosen responses in the labeled_tp stratum were judged clinically correct.
  • standard math The marginal distribution decomposes as PX = alpha Pp + (1 - alpha) Pn.
    Standard mixture identity used in Proposition A.1 to derive the unbiased PU loss.
  • domain assumption Per-finding preferences compose into a global report optimum through a monotone single-action-improvement path.
    Appendix A.3 relies on this to argue that local contrastive pairs recover the globally optimal report. The authors themselves call the set-based formulation an analytical abstraction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Positive-Unlabeled Preference Optimization For Chest X-ray Report Generation." pith.science (2026). https://pith.science/paper/7ZZLDDBU

@misc{pith2026260805341,
  author       = {Pith},
  title        = {Pith review of: Positive-Unlabeled Preference Optimization For Chest X-ray Report Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7ZZLDDBU}},
  note         = {Machine review of arXiv:2608.05341}
}
read the original abstract

Vision-Language Models (VLMs) for radiology report generation are typically trained on retrospective clinical reports, which suffer from omission noise: clinically present findings are left unreported due to the omission of subtle findings. For example, prior studies show that cardiomegaly may be omitted from ICU chest X-ray reports when the imaging request is focused on monitoring support device placement. As a result, models trained with standard approaches inherit these omissions, learning to under-report findings themselves. We propose PU-DPO, a preference optimization framework to prevent omission noise from corrupting the preference signal. We reformulate the objective under a positive-unlabeled (PU) learning framework, treating absent mentions as unlabeled rather than truly negative. Our framework provides preference supervision using constructed contrastive pairs, generated using edits to model responses, producing variants that explicitly mention or omit a specific finding. Generated responses that mention the finding are naturally preferred in the context of visual evidence. Across semi-synthetic experiments and analyses on real-world chest radiograph benchmarks where adjudicated labels are available, PU-DPO yields consistent gains in detection rates and recovery of hidden positives across multiple pathologies, and is more robust to omission noise than prior approaches.

Figures

Figures reproduced from arXiv: 2608.05341 by the authors.

Figure 1
Figure 1. Overview of Omission Noise and Proposed PU Learning Framework. (Left) We highlight the challenge of omission noise in training medical VLMs, where clinical findings are visually present but omitted from reports due to standard clinical workflow variability. (Right) Our proposed approach mitigates this by constructing preference pairs that evaluate the presence versus absence of a single finding. We then apply a nois… view at source ↗
Figure 2
Figure 2. Our PU learning approach maintains a high F1 score as the noise rate increases. Sensitivity [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Relative improvement in Micro-F1, omission recovery, and specificity compared to the [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Relative improvement in Micro-F1, sensitivity, and specificity compared to the base model [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Certain pathologies (atelectasis) and presentation subtypes (small unilateral pleural ef [PITH_FULL_IMAGE:figures/full_fig_p028_5.png]
Figure 6
Figure 6. Figure 6: Subgroup-wise marginal sensitivity for the NV-reason model, showing consistent patterns [PITH_FULL_IMAGE:figures/full_fig_p028_6.png]
Figure 7
Figure 7. Figure 7: We find consistent patterns with the GPT-OSS 120B model extracting clinical labels from [PITH_FULL_IMAGE:figures/full_fig_p028_7.png]
Figure 8
Figure 8. Figure 8: We find consistent patterns in the MIMIC dataset. [PITH_FULL_IMAGE:figures/full_fig_p029_8.png]
Figure 9
Figure 9. Figure 9: We plot the TPR of pleural effusion for the top 20 most frequent intersectional subtypes [PITH_FULL_IMAGE:figures/full_fig_p029_9.png]
Figure 10
Figure 10. Figure 10: We plot the TPR of atelectasis for the top 20 most frequent intersectional subtypes defined [PITH_FULL_IMAGE:figures/full_fig_p030_10.png]
Figure 11
Figure 11. Figure 11: Calibration of learned preference probabilities for pleural effusion. PU-DPO leads to [PITH_FULL_IMAGE:figures/full_fig_p032_11.png]
Figure 12
Figure 12. Figure 12: Pathology-specific results for F1, omission recovery, and specificity, Medgemma model. [PITH_FULL_IMAGE:figures/full_fig_p036_12.png]
Figure 13
Figure 13. Figure 13: Pathology-specific results for F1, omission recovery, and specificity, NV-reason model. [PITH_FULL_IMAGE:figures/full_fig_p037_13.png]
Figure 14
Figure 14. Figure 14: Pathology-specific results for F1, omission recovery, and specificity, Llava-RAD model. [PITH_FULL_IMAGE:figures/full_fig_p037_14.png]
Figure 15
Figure 15. Figure 15: Sensitivity to α misspecification on F1, omission recovery, and specificity, Medgemma model. 37 [PITH_FULL_IMAGE:figures/full_fig_p037_15.png]
Figure 16
Figure 16. Figure 16: Sensitivity to α misspecification on F1, omission recovery, and specificity, NV-reason model [PITH_FULL_IMAGE:figures/full_fig_p038_16.png]
Figure 17
Figure 17. Figure 17: Sensitivity to α misspecification on F1, omission recovery, and specificity, Llava-RAD model. 38 [PITH_FULL_IMAGE:figures/full_fig_p038_17.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 31 canonical work pages

  1. [1]

    Scaling medical imaging report generation with multimodal reinforcement learning

    Scaling medical imaging report generation with multimodal reinforcement learning , author=. arXiv preprint arXiv:2601.17151 , year=

    work page Pith review arXiv

  2. [2]

    arXiv preprint arXiv:2504.20930 , year=

    Chestx-reasoner: Advancing radiology foundation models with reasoning through step-by-step verification , author=. arXiv preprint arXiv:2504.20930 , year=

    arXiv

  3. [3]

    Advances in neural information processing systems , volume=

    Co-teaching: Robust training of deep neural networks with extremely noisy labels , author=. Advances in neural information processing systems , volume=

  4. [4]

    Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

    Learning classifiers from only positive and unlabeled data , author=. Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

    work page

  5. [5]

    arXiv preprint arXiv:2412.14922 , year=

    Robustft: Robust supervised fine-tuning for large language models under noisy response , author=. arXiv preprint arXiv:2412.14922 , year=

    arXiv

  6. [6]

    Applied Sciences , volume=

    Noise-Aware Direct Preference Optimization for RLAIF , author=. Applied Sciences , volume=. 2025 , publisher=

    work page 2025

  7. [7]

    arXiv preprint arXiv:2404.04102 , year=

    ROPO: Robust Preference Optimization for Large Language Models , author=. arXiv preprint arXiv:2404.04102 , year=

    arXiv

  8. [8]

    arXiv preprint arXiv:2403.00409 , year=

    Provably robust dpo: Aligning language models with noisy feedback , author=. arXiv preprint arXiv:2403.00409 , year=

    arXiv

Show all 60 references
  1. [9]

    Radiology: Artificial Intelligence , volume=

    Anatomy-specific progression classification in chest radiographs via weakly supervised learning , author=. Radiology: Artificial Intelligence , volume=. 2024 , publisher=

    2024

  2. [10]

    Advances in Neural Information Processing Systems , volume=

    Mixture proportion estimation and PU learning: A modern approach , author=. Advances in Neural Information Processing Systems , volume=

  3. [11]

    Journal of Machine Learning Research , volume=

    Risk bounds for positive-unlabeled learning under the selected at random assumption , author=. Journal of Machine Learning Research , volume=

  4. [12]

    Advances in neural information processing systems , volume=

    Positive-unlabeled learning with non-negative risk estimator , author=. Advances in neural information processing systems , volume=

  5. [13]

    Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP) , pages=

    Combining automatic labelers and expert annotations for accurate radiology report labeling using BERT , author=. Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP) , pages=

    2020

  6. [14]

    Nature Machine Intelligence , volume=

    Benchmarking saliency methods for chest X-ray interpretation , author=. Nature Machine Intelligence , volume=. 2022 , publisher=

    2022

  7. [15]

    Advances in neural information processing systems , volume=

    Aligning llm agents by learning latent preference from user edits , author=. Advances in neural information processing systems , volume=

  8. [16]

    arXiv preprint arXiv:2312.11456 , year=

    Iterative preference learning from human feedback: Bridging theory and practice for rlhf under kl-constraint , author=. arXiv preprint arXiv:2312.11456 , year=

    arXiv

  9. [17]

    arXiv preprint arXiv:2404.14367 , year=

    Preference fine-tuning of llms should leverage suboptimal, on-policy data , author=. arXiv preprint arXiv:2404.14367 , year=

    arXiv

  10. [18]

    arXiv preprint arXiv:2410.06961 , year=

    Self-boosting large language models with synthetic preference data , author=. arXiv preprint arXiv:2410.06961 , year=

    arXiv

  11. [19]

    Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

    Self-instruct: Aligning language models with self-generated instructions , author=. Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

  12. [20]

    Editgrpo: Reinforcement learning with post-rollout edits for clinically accurate chest x-ray report generation , author=. Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association...

  13. [21]

    Advances in neural information processing systems , volume=

    Training language models to follow instructions with human feedback , author=. Advances in neural information processing systems , volume=

  14. [22]

    Proceedings of the Conference on Health, Inference, and Learning , pages=

    VisualCheXbert: addressing the discrepancy between radiology report labels and image labels , author=. Proceedings of the Conference on Health, Inference, and Learning , pages=

  15. [23]

    Radiology , volume=

    Chest radiograph interpretation with deep learning models: assessment with radiologist-adjudicated reference standards and population-adjusted evaluation , author=. Radiology , volume=. 2020 , publisher=

    2020

  16. [24]

    IEEE transactions on neural networks and learning systems , volume=

    Classification in the presence of label noise: a survey , author=. IEEE transactions on neural networks and learning systems , volume=. 2013 , publisher=

    2013

  17. [25]

    International conference on algorithmic learning theory , pages=

    Positive and unlabeled examples help learning , author=. International conference on algorithmic learning theory , pages=. 1999 , organization=

    1999

  18. [26]

    Advances in neural information processing systems , volume=

    Theoretical comparisons of positive-unlabeled learning against positive-negative learning , author=. Advances in neural information processing systems , volume=

  19. [27]

    IEICE TRANSACTIONS on Information and Systems , volume=

    Class prior estimation from positive and unlabeled data , author=. IEICE TRANSACTIONS on Information and Systems , volume=. 2014 , publisher=

    2014

  20. [28]

    arXiv preprint arXiv:1601.01944 , year=

    Nonparametric semi-supervised learning of class proportions , author=. arXiv preprint arXiv:1601.01944 , year=

    arXiv

  21. [29]

    International conference on learning representations , year=

    Learning from positive and unlabeled data with a selection bias , author=. International conference on learning representations , year=

  22. [30]

    Second International Workshop on Learning with Imbalanced Domains: Theory and Applications , pages=

    Learning from positive and unlabeled data under the selected at random assumption , author=. Second International Workshop on Learning with Imbalanced Domains: Theory and Applications , pages=. 2018 , organization=

    2018

  23. [31]

    Diagnostics , volume=

    Interpretation of thoracic radiography shows large discrepancies depending on the qualification of the physician—quantitative evaluation of interobserver agreement in a representative emergency department scenario , author=. Diagnostics , volume=. 2021 , publisher=

    2021

  24. [32]

    International conference on machine learning , pages=

    Convex formulation for learning from positive and unlabeled data , author=. International conference on machine learning , pages=. 2015 , organization=

    2015

  25. [33]

    Advances in Neural Information Processing Systems , volume=

    A variational approach for learning from positive and unlabeled data , author=. Advances in Neural Information Processing Systems , volume=

  26. [34]

    Machine learning , volume=

    Learning from positive and unlabeled data: A survey , author=. Machine learning , volume=. 2020 , publisher=

    2020

  27. [35]

    arXiv preprint arXiv:2507.05201 , year=

    Medgemma technical report , author=. arXiv preprint arXiv:2507.05201 , year=

    arXiv

  28. [36]

    arXiv preprint arXiv:2407.07880 , year=

    Towards robust alignment of language models: Distributionally robustifying direct preference optimization , author=. arXiv preprint arXiv:2407.07880 , year=

    arXiv

  29. [37]

    Advances in neural information processing systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in neural information processing systems , volume=

  30. [38]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    Anatomy-guided weakly-supervised abnormality localization in chest x-rays , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=. 2022 , organization=

    2022

  31. [39]

    Advances in Neural Information Processing Systems , volume=

    Perplexity-aware correction for robust alignment with noisy preferences , author=. Advances in Neural Information Processing Systems , volume=

  32. [40]

    arXiv preprint arXiv:2510.01458 , year=

    How Well Can Preference Optimization Generalize Under Noisy Feedback? , author=. arXiv preprint arXiv:2510.01458 , year=

  33. [41]

    arXiv preprint arXiv:2510.23968 , year=

    Reasoning Visual Language Model for Chest X-Ray Analysis , author=. arXiv preprint arXiv:2510.23968 , year=

  34. [42]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages=

    Chexpo: Preference optimization for chest x-ray vlms with counterfactual rationale , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=

  35. [43]

    CURV: Coherent Uncertainty-Aware Reasoning in Vision-Language Models for X-Ray Report Generation , author=

  36. [44]

    Nature Communications , volume=

    A clinically accessible small multimodal radiology model and evaluation metric for chest X-ray findings , author=. Nature Communications , volume=. 2025 , publisher=

    2025

  37. [45]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. Iclr , volume=

  38. [46]

    Proceedings of Machine Learning Research , volume=

    RadVLM-GRPO: Enhancing Chest X-ray Report Generation and Visual Grounding via Reinforcement Learning , author=. Proceedings of Machine Learning Research , volume=

  39. [47]

    Patterns , volume=

    Evaluating progress in automatic chest x-ray radiology report generation , author=. Patterns , volume=. 2023 , publisher=

    2023

  40. [48]

    Nature Medicine , volume=

    Collaboration between clinicians and vision--language models in radiology report generation , author=. Nature Medicine , volume=. 2025 , publisher=

    2025

  41. [49]

    arXiv preprint arXiv:1901.07042 , year=

    MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographs , author=. arXiv preprint arXiv:1901.07042 , year=

    1901 arXiv

  42. [50]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  43. [51]

    arXiv preprint arXiv:2405.19538 , year=

    Chexpert plus: Augmenting a large chest x-ray dataset with text radiology reports, patient demographics and additional image formats , author=. arXiv preprint arXiv:2405.19538 , year=

    arXiv

  44. [52]

    Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

    Bleu: a method for automatic evaluation of machine translation , author=. Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

  45. [53]

    Proceedings of the third conference on machine translation: Research papers , pages=

    A call for clarity in reporting BLEU scores , author=. Proceedings of the third conference on machine translation: Research papers , pages=

  46. [54]

    Text summarization branches out , pages=

    Rouge: A package for automatic evaluation of summaries , author=. Text summarization branches out , pages=

  47. [55]

    arXiv preprint arXiv:2106.14463 , year=

    Radgraph: Extracting clinical entities and relations from radiology reports , author=. arXiv preprint arXiv:2106.14463 , year=

    arXiv

  48. [56]

    arXiv preprint arXiv:1904.09675 , year=

    Bertscore: Evaluating text generation with bert , author=. arXiv preprint arXiv:1904.09675 , year=

    1904 arXiv

  49. [57]

    arXiv preprint arXiv:2512.16145 , year=

    MRG-R1: Reinforcement Learning for Clinically Aligned Medical Report Generation , author=. arXiv preprint arXiv:2512.16145 , year=

  50. [58]

    the method of paired comparisons , author=

    Rank analysis of incomplete block designs: I. the method of paired comparisons , author=. Biometrika , volume=. 1952 , publisher=

    1952

  51. [59]

    JMIR Formative Research , volume=

    Vision-language model for generating textual descriptions from clinical images: model development and validation study , author=. JMIR Formative Research , volume=. 2024 , publisher=

    2024

  52. [60]

    arXiv preprint arXiv:2512.16201 , year=

    Visual Alignment of Medical Vision-Language Models for Grounded Radiology Report Generation , author=. arXiv preprint arXiv:2512.16201 , year=

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.